Privacy Preserved Brain Disorder Diagnosis Using Federated Learning

1  Introduction

The use of wearable healthcare devices has increased in popularity as technology has advanced, and users have become more interested in tracking and monitoring their health. People’s daily activities and health are closely intertwined [1,2]. People can better understand their health status by tracking their activities through wearable technology, such as smart glasses, smartphones, and wristbands [3–5]. These devices can track various health metrics, from physical activity and sleep patterns to heart rate and blood glucose levels. They can also provide alerts and reminders, enabling users to take action if a potential health issue arises. These devices’ increasing availability and affordability have led to a growing interest in using wearable healthcare technology to promote health and prevent disease. Early detection of numerous cognitive illnesses, including Parkinson’s disease, tumors [4,6–11], and minor vascular diseases, may be possible using wearable healthcare [12]. Additional uses include fall detection, sports monitoring, and mental health assessment [13–15].

Parkinson’s disease is a chronic neurodegenerative disorder brought about through the slow death of dopaminergic brain cells [16–18]. Early Parkinson’s disease detection, diagnosis, and therapy may slow or cease the disease’s progression and lengthen the sufferers’ lives [19–22]. However, people are typically unaware of the early signs of Parkinson’s disease in their daily lives because the initial phases of Parkinson’s only show mild or undetectable non-motor symptoms (such as voice abnormalities, mood disorders, and sleep disturbances) [23–26]. The solutions based on information and communication technologies have presented various methods to measure the motor condition of Parkinson’s patients. Human-mobile interaction has become a common phenomenon in recent years, creating massive data streams [27]. The most extensive smartphone-based Parkinson’s disease study was conducted using designed sensors on mobile devices to capture and extract data about Parkinson’s disease [28,29], which engaged participants remotely, asked them to self-report their statistical profiles and conducted standardized tests to obtain data over 6 months.

Machine learning models are routinely developed through user data to track the user’s health state for intelligent healthcare [30,31]. Despite the quantity of data in various organizations, sharing the data is not practical owing to security and privacy considerations [32–34]. Because of this, it is challenging to build effective models using valuable data. Personalizing is the other significant issue. Most approaches are based on a standard service model for all users. The machine learning model is disseminated to the connected systems to monitor the health data regularly after receiving enough data to train it [35,36]. This procedure deficit is personalizing. It is observed that each user has a unique set of physical traits and everyday routines. The ordinary model cannot provide security and privacy to personalized healthcare data. Providing data security and privacy is one of several significant difficulties facing wearable healthcare devices [37]. Sufficient user data is frequently employed to train machine learning models for healthcare applications and to track health status [38–41]. Recently, federated learning has drawn considerable interest as a novel method that enables deep and machine learning algorithms to leverage global learning among many people while protecting user data safety [42–46].

This paper makes the following major contributions:

•   Propose an approach to detecting and diagnosing brain disorders using federated learning by extracting the data from the Kay Elemetrics voice-disordered database and perform data preprocessing to remove inconsistencies and irrelevant values.

•   The proposed model trains the local model first and sends the model’s hyperparameters to the server. Then, the server sum up the updated parameters and trains the model by keeping the privacy of each client’s data.

•   The experiment shows that the federated learning model based DNN model achieves an accuracy score of 82.82% for brain disorders detection compared to conventional methods and preserves client data protection.

Section 2 discusses the related work for diagnosing Parkinson’s disease. Section 3 discusses the proposed framework based on federated learning. Section 4 explains the experimental results and provides a discussion. Section 5 concludes this paper and provides future research.

2  Literature Review

This section provides the history of Parkinson’s disease using machine learning and deep learning techniques. The background of the healthcare devices that use federated learning is also provided.

2.1 Machine Learning Techniques

The earlier diagnosis and prognosis of various ailments depended on AI, including Huntington’s illness [47], identifying speech disorders [48], and vocal evaluation influenced by bio-system [49]. Author [50] identified non-speech body movements as the new Parkinson’s disease biomarker and leveraged smartphone usage data to monitor passive Parkinson’s disease in daily life without disturbing the user. The PDVocal, a smartphone-based, end-to-end, privacy-preserving technology for early Parkinson’s disease detection, is presented. During routine phone calls, PDVocal can passively detect Parkinson’s disease digital biomarkers in the audio data. The long-term tracking protocol and opportunistic learning knob methods are investigated to achieve a reliable Parkinson’s disease detection system. The proposed system is evaluated based on real-world interactions from publicly accessible data sources and a dataset compiled from 890 participants. The proposed system achieves 98.8% accuracy and 1.0% FRR with a 0.5-s segment size; with a 1-s segment size, it achieves 99% accuracy and 1.1% FRR. The author [51] proposed using the machine learning technique for Parkinson’s disease detection using smartphone applications. The study observes the five tasks completed remotely: reaction time, voice, balancing, tapping of a finger, and gait. The data regarding Parkinson’s disease is obtained from 6,148 smartphone activity assessments completed by 129 individuals. The proposed model increased the standard deviation by 16.3 (5.6) points in response to dopaminergic medication.

2.2 Deep Learning Techniques

Author [52] proposes a deep neural network algorithm on the Parkinson’s Telemonitoring Voice Dataset of patients from the UCI repository. The TensorFlow library and Python language are applied to the proposed deep neural network approach. The dataset constitutes 16 attributes used in this research and two severity scores (motor UPDRS and total UPDRS) for the prediction. The proposed approach was applied to both scores individually. The motor UPDRS score achieves better accuracy than the total UPDRS score: 83.367%. Author [53] implements a machine learning-based technique for classifying Parkinson’s disease by employing a dataset from smartphone-based walking, speaking, typing, and memory tests. The proposed approach consists of two stages of performing individual training. The first consists of customized models to evaluate the intensity of symptoms in cases where only one test is performed. The second is an EAM (Evidence Aggregation Model) to combine all single-test assessments into a complete diagnostic score. The proposed approach significantly outperforms the baselines at predicting by using data from a cohort of 1,853 participants. The highest AUC-ROC achieved is 0.85.

Authors in [54] explain that multi-source ensemble learning, a technique that combines dataset deconstructing with ensemble learning, obtains a participant retention rate of 100% and allows participants with incomplete information to be utilized within the training of machine learning models. According to the proposed approach, 91.2% of a cohort of 1,513 participants provided incomplete data on tapping, walking, voice, or memory tests. Convolutional neural networks that use the volume of data available and multi-source ensemble learning have been shown to boost Parkinson’s dataset classification accuracy from 73.1% to 82.0% by comparing existing methods. It was discovered that using multi-channel convolutional neural networks (CNNs) and creating models employing a large cohort of participants contributed to the accuracy improvement. Authors in [55] proposed the utilization of DeepVoice, a program developed for PD diagnosis based on voiceprints and merging deep learning and mobile health. The proposed approach evaluates input data type, voice length, and a neural network model on a massive Parkinson’s dataset. According to the result, DeepVoice can identify the Parkinson’s dataset with an accuracy of 90.45 ± 1.71% using an audio clip 10 s long. Table 1 presents the summary of existing works.

2.3 Federated Learning Techniques

The author [56] proposed a privacy-based, federated learning approach to protecting patients’ data [56]. The BraTS 2018 dataset builds and analyzes efficient federated learning techniques for segmenting brain tumors. The dataset includes 285 individuals with brain tumors recognized through multi-parametric pre-operative MRI scans. Authors in [57] proposed a machine learning-cum-attention mechanism to segment brain tumors effectively. Akter et al. [41] introduce a framework with Distributed Protection (DP) to secure the healthcare sector using IoT technologies [41]. The datasets utilized in this study are from the MNIST database, CIFAR10, and the COVID-19 dataset and are used to show the proposed technique’s performance from a healthcare perspective. An iteration-based converged CNN at Edge Layer is proposed to reconcile federated learning’s privacy protection and algorithm performance across an IoT network. After several iterations, the findings show the highest accuracy rate of 90%, outperforming other baseline techniques. Additionally, the proposed approach satisfies the privacy and security concept more quickly and effectively [41]. Authors in [58] presented the framework of MetisFL Federated Learning, a flexible architecture enabling a range of communication channels with robust security and privacy methods. It established its effectiveness in difficult cognitively and statistically varied contexts for neuroimaging applications, the prediction of Alzheimer’s disease, and Brain-AGE. The findings show the best accuracy is 0. 8633 ± 0.0013%, precision is 0.8098 ± 0.0043%, recall is 0.8132 ± 0.0097%, f1-measure is 0.8114 ± 0.003%, AUC PR is 0.8682 ± 0.0009%, and AUC-ROC is 0. 8971 ± 0.0006% on the federated 5AOB model.

In conclusion, various machine and deep learning methods have been investigated to develop Parkinson’s disease prediction methods. However, they do not offer reassuring evidence of improved accuracy and do not prioritize data protection. The federated learning research previously mentioned centered on resolving various healthcare challenges [56,58]. To preserve data privacy that ML and DL models do not account for, we proposed a federated learning framework in this study that is built on a DNN model.

3  Proposed Methodology

This section represents the idea of federated learning, model architecture, and data preprocessing since they serve the basics of the proposed technique. Fig. 1 depicts the steps performed in the proposed methodology, which consists of three crucial steps. Training initializing is the first step. Based on the target application, the federated learning server, a cloud server, evaluates the significant types of data and trains the hyper-parameters, such as the learning rate, the number of epochs, the activation function, and the Adam optimizer.

Figure 1: Overview of proposed deep federated learning framework

The neural network DNN receives client specifications and multiple hyper-parameters. The next level is to train the DNN model. Every client starts to gather new data and change the parameter on the local model (Myx) based on the global model (Gyx), wherein y is the index for the following iteration. Every client is looking for the optimal situation to decrease the loss. Ultimately, the modified parameters are submitted to the federated learning server regularly. The aggregation of the global model is the third level. At this level, the findings from various clients on the server side are combined, and then the modified parameters are delivered to every client. The server lowers the global loss function by following Eq. (1):

Loss(Gy )=1m∑x=mx=1Loss(Myx)(1)

The process is repeated until good results are achieved or the loss is continuously reduced. The proposed model’s steps are provided in Algorithm 1. The Parkinson’s disease dataset is taken as input. Exploratory Data Analysis (EDA) is performed to remove the inconsistencies. Before prepossessing, the dataset is split into two windows for clients 1 and 2. Data preprocessing is performed using Synthetic Minority Over-Sampling Technique (SMOTE) and normalization to balance the dataset. The testing and training portions of the dataset are divided into 25% and 75%, respectively, after prepossessing. The global model is created along with two local models where all models run for three rounds. The loss is calculated through binary cross-entropy and evaluation metrics. After obtaining the results, aggregate both client results and the overall results.

3.1 Dataset Selection, Preprocessing, and Feature Extraction

Next, the EDA is performed on a dataset extracted from the database to eliminate inconsistencies and irrelevant values. This study uses the dataset extracted from the database of Kay Elemetrics voice disordered [59]; the dataset is available at Kaggle. The database has 707 standard and abnormal voice samples from various individuals who suffer traumatic, biological, and neurological disorders. These are samples of the three forms of strange voice and speech signals, and 53 samples of ordinary people are used as controls. In the database, each voice sample is an average of 2 s long, 16-bit raw PCM audio and was collected in controlled, quiet acoustic settings. Several voice samples were down-sampled using anti-aliasing from 50 to 25 kHz.

The EDA is utilized as a strategy for learning from the data instead of using it for formal modeling or hypothesis testing. To summarize the statistical parameters of the datasets, it is helpful to conduct an analysis focusing on four crucial factors: the spread measurements, central tendency measurements, the outlier’s existence, and the disturbance’s shape. The dataset contains binary classes; 1 indicates the patient has Parkinson’s disease, and 0 indicates a healthy person. The total training data is 198. The dataset is divided into two dormers because this study works with two clients, each with its data. Client 1 has 98 data, and Client 2 has 97 data.

Two preprocessing methods, SMOTE and normalization, are applied to the dataset to normalize and balance the data. SMOTE is an oversampling method that uses fictitious samples to represent the minority classes. This technique helps to overcome the overfitting problem that results from random oversampling.

Client 1 has 154 training data after applying the SMOTE to the dataset, while Client 2 has 140. The scaling of input attributes is essential for a specific model. The min-max scaling is used for feature normalization that helps to structure the low variance and ambiguous dataset and maintain the data integrity. Eq. (2) is used for data normalization.

Znorm=Zi−ZminZmax−Zmin(2)

3.2 Model Architecture

The number of neurons and layers is crucial when modeling neural network topology. Most of the time, signals are only transferred in one direction from input to output within the network. Since each neuron’s output does not affect the neuron, there is no loop. The DNN is the neural network architecture that is most well-known and is widely utilized to process the data, define the input and output neurons. The DNN algorithm usually consists of three layers: an input layer, a different hidden layer (dropout, dense, etc.), and the output layer. The architecture of the DNN model is given in Fig. 2.

Figure 2: Proposed architecture of DNN model

4  Results and Analysis

The experiment is on the dataset extracted from the Kay Elemetrics voice-disordered database. In the experiment, one server and two clients are engaged. The two clients use the Parkinson’s diagnosis dataset to train the DNN model. The DNN model is trained using the given dataset, and the random weight is initialized. The specifics of the DNN model are given in the section on model architecture. We first gathered 195 dataset sample samples, after which we performed an EDA on the dataset to eliminate the inconsistencies and irrelevant values. The SMOTE method balances the dataset after the EDA has been applied. The total training data after the SMOTE is 294 in total. The dataset is divided into two parts: 25% is used for testing, and 75% is used to train the model. The values are all fitted into the range [0, 1] using the min-max normalization procedure. The experiment is conducted in two classes. Each client’s results are reviewed three times for loss prevention. The label encoder method converts the label into a machine-readable format. At the server end, the client’s results are combined. The experiment uses different evaluation metrics: precision, accuracy, f1-measure, and recall. After server-side aggregation, the DNN model obtains 82.82% accuracy.

4.1 Server-Based Training with Log Data

One server is the main parameter, and two clients comprise a federated learning system. The server decides which client/node is used at an initial level of the model training stage and collects any modifications received. Data logged on servers is used to train the DNN model. Before training, the logs are anonymized and cleared of personal identifying details. During the fit-round, the clients transmit the output to the server side. Both clients communicate the test result to the server during the evaluation round, which combines the results. The DNN achieves the best accuracy of 82.82% on the server side.

4.2 Federated Training with Client 1

Client 1 conducts the experiment using a sequential DNN model with two dense and three dropout layers as hidden layers. The DNN model runs three test rounds. The experiment conducts three rounds of evaluation metrics: precision, accuracy, f1-measure, and recall. Table 2 represents the experimental results of client 1. The study runs the DNN model for round 1, and the result obtains accuracy, recall, precision, and F1 measure. In the first round, the DNN model provides 91.0% accuracy, 88.0% precision, 94.0% recall, and 99.0% F1-measure. Analyze the results to prevent the model from overfitting once again. In round 2, the DNN model achieved an accuracy of 77.0%, a precision of 70.0%, a recall of 88.0%, and a 78.0% F1-measure. In round 3, we obtained an accuracy of 74.0%, a precision of 67.0%, a recall of 88.0%, and a 76.0% F1-measure. From client one, round 1 provides the highest result.

In Fig. 3, the highest result is visualized. In Fig. 3a, it can be noticed that the training accuracy started increasing from 0.5% and kept increasing continuously to around 0.95%. Similarly, validation accuracy started increasing from 0.6% and kept increasing continuously to around 0.96%. The graph in Fig. 3b depicts the training and validation loss; in the training process, the loss is reduced on each epoch, which increases the model’s performance. The training loss started decreasing from 0.7% and kept decreasing continuously to around 0.2%. Similarly, validation accuracy started decreasing from 0.65% and kept decreasing continuously to around 0.2%. The graph in Fig. 3c indicates the ROC. The ROC curve represents better performance near the corner of the top-left.

Figure 3: Graphs of client 1 best results

4.3 Federated Training with Client 2

Client 2 conducts the experiment using a sequential DNN model. The DNN algorithm has two dense and three dropout layers as hidden layers. The client conducted two experiments three times and returned the results in the evaluation and fit rounds. The DNN model runs three test rounds. The experiment conducts three rounds of evaluation metrics: precision, accuracy, f1-measure, and recall.

Table 3 represents the experimental results of client 2. DNN achieves an accuracy of 95.0%, 96.0% precision, 96.0% recall, and 96.0% F1-measure. In round 2, the DNN model achieved an accuracy of 95.0%, 96.0% precision, 96.0% recall, and 96.0% F1-measure. Round 3 obtained an accuracy of 95.0%, 96.0% precision, 96.0% recall, and 96.0% F1-measure. From client two, round 3 provides the highest result.

Fig. 4 depicts the highest result. Fig. 4a represents the accuracy graph in round 3 of client 2 obtained the highest validation accuracy. It can be noticed that the training accuracy started increasing from 0.77% and kept increasing continuously to around 0.82%. Similarly, validation accuracy started from 0.82% and kept decreasing continuously to around 0.74%. In Fig. 4b, the graph shows that the loss is reduced on each epoch, which increases the model’s performance. The training loss started decreasing from 0.48% and kept increasing continuously to around 0.46%. Similarly, validation accuracy started decreasing from 0.4% and kept decreasing continuously to around 0.2%. The third graph indicates the Receiver Operating Characteristic (ROC) in Fig. 4c. The ROC curve represents better performance near the corner of the top-left.

Figure 4: Graphs of client 2 best results (Round 3)

In Fig. 5a, the proposed approach’s confusion matrix of client 1 is represented graphically and provides an overview of how a classification algorithm operates. Because it has more continuous and better true positive and negative values and fewer false negative and false positive values, the proposed method performs better. Fig. 5b presents the proposed approach's confusion matrix for client 2. It provides an overview of how a classification algorithm operates. Because it has more continuous, better true positive and negative values and fewer false negative and false positive values, the proposed method performs better.

Figure 5: Proposed approach’s confusion matrix for Parkinson’s disease detection

4.4 Limitations and Discussions

This research suggests FL-based models be used to identify and diagnose Parkinson’s disease as a solution to this problem. The model employed a dataset extracted from the Kay Elemetrics voice disorder database; however, the extracted dataset was small. It is challenging, therefore, for single sites with tiny, labeled datasets to create unique AI systems for screening patients. This is because the patient population on which the model is built is often minuscule, which, in turn, significantly reduces the model’s generalizability. FL eliminates this hurdle. By collaborating, many organizations can combine their data to develop a more accurate global model for a broader range of patients. Since FL prioritizes the protection of patient data, there is no explicit data sharing in this cooperative process. Although FL makes it possible for more cooperative ML, it also presents a particular variety of challenges, such as the preceding: in FL networks, where the data generated by each device is maintained locally, communication is a significant barrier. Another challenge is that not all ML applications are compatible with FL. The developer must discover alternative solutions if the model is too large to execute on consumer devices to protect user privacy, as this study utilizes a small dataset that positively impacts model performance.

5  Conclusion

This paper proposed a federated learning based DNN model to detect and diagnose Parkinson’s disease. The novelty of the proposed approach lies in providing privacy to the user data. For experimentation, we extracted the data from the Kay Elemetrics voice-disordered database. The dataset is divided into two clients to create the training model because this study works with two clients, each with its data and one server. Two clients were used to train the DNN model, and the server end was combined with an accuracy of 82.82%. To lessen the over-fitting aspect, every client reviewed the outcomes three times. The best result of round 1 of client one was obtained with a 91.0% accuracy, and client 2 of round 1 obtained the best result with a 95.0% accuracy. The DNN model’s ROC curve average of 99.3% demonstrates the suggested method’s excellent performance on the tested dataset. The findings showed that federated learning accurately secures the privacy of client data. According to a system efficiency study, side training times and storage costs favor medical devices with constrained resources. In the future, we intend to research training FL-model on heterogeneous devices to improve the Parkinson’s disease detection process.

Acknowledgement: This work was supported by the Deanship of Scientific Research at Prince Sattam bin Aziz University under the Research Project (PSAU/2023/01/22425).

Funding Statement: This work was supported by the Deanship of Scientific Research at Prince Sattam bin Aziz University under the Research Project (PSAU/2023/01/22425).

Conflicts of Interest: The authors declare they have no conflicts of interest to report regarding the present study.

References

1. A. R. Javed, L. G. Fahad, A. A. Farhan, S. Abbas, G. Srivastava et al., “Automated cognitive health assessment in smart homes using machine learning,” Sustainable Cities and Society, vol. 65, no. 9, pp. 102572, 2021. [Google Scholar]

2. A. R. Javed, M. U. Sarwar, M. O. Beg, M. Asim, T. Baker et al., “A collaborative healthcare framework for shared healthcare plan with ambient intelligence,” Human-centric Computing and Information Sciences, vol. 10, no. 1, pp. 1–21, 2020. [Google Scholar]

3. A. R. Javed, R. Faheem, M. Asim, T. Baker and M. O. Beg, “A smartphone sensors-based personalized human activity recognition system for sustainable smart cities,” Sustainable Cities and Society, vol. 71, no. 5, pp. 102970, 2021. [Google Scholar]

4. S. Alsubai, H. U. Khan, A. Alqahtani, M. Sha, S. Abbas et al., “Ensemble deep learning for brain tumor detection,” Frontiers in Computational Sciences, vol. 16, pp. 788347, 2022. [Google Scholar]

5. S. Alsubai, A. Alqahtani, M. Sha, S. Abbas, A. Almadhor et al., “Smart home-based complex interwoven activities for cognitive health assessment,” Journal of Sensors, vol. 2022, no. 2, pp. 1–10, 2022. [Google Scholar]

6. H. Mughal, A. R. Javed, M. Rizwan, A. S. Almadhor and N. Kryvinska, “Parkinsonâs disease management via wearable sensors: A systematic review,” IEEE Access, vol. 10, pp. 35219–35237, 2022. [Google Scholar]

7. A. Mohiyuddin, A. Basharat, U. Ghani, V. Peter, S. Abbas et al., “Breast tumor detection and classification in mammogram images using modified yolov5 network,” Computational and Mathematical Methods in Medicine, vol. 2022, no. 3, pp. 1–16, 2022. [Google Scholar]

8. J. Zhang, “Mining imaging and clinical data with machine learning approaches for the diagnosis and early detection of parkinson’s disease,” npj Parkinson’s Disease, vol. 8, no. 1, pp. 1–15, 2022. [Google Scholar]

9. M. A. Chalah and S. S. Ayache, “Noninvasive brain stimulation and psychotherapy in anxiety and depressive disorders: A viewpoint,” Brain Sciences, vol. 9, no. 4, pp. 82, 2019. [Google Scholar] [PubMed]

10. R. Abou Khzam, N. D. El Jalbout, R. Seif and A. Sadaka, “An unusual presentation of convergence insufficiency in a patient with Parkinson’s disease stimulated by deep brain stimulation,” American Journal of Ophthalmology Case Reports, vol. 26, pp. 101531, 2022. [Google Scholar] [PubMed]

11. V. Matanova, Z. Kostova and M. Kolev, “Brain-based treatment new approach or a well-forgotten old one?” Journal of Evaluation in Clinical Practice, vol. 24, no. 4, pp. 859–863, 2018. [Google Scholar] [PubMed]

12. Y. Chen, C. Hu, B. Hu, L. Hu, H. Yu et al., “Inferring cognitive wellness from motor patterns,” IEEE Transactions on Knowledge and Data Engineering, vol. 30, no. 12, pp. 2340–2353, 2018. [Google Scholar]

13. A. Alqahtani, S. Alsubai, M. Sha, V. Peter, A. S. Almadhor et al., “Falling and drowning detection framework using smartphone sensors,” Computational intelligence and neuroscience, vol. 2022, pp. 6468870, 2022. [Google Scholar] [PubMed]

14. J. Wang, Y. Chen, S. Hao, X. Peng and L. Hu, “Deep learning for sensor-based activity recognition: A survey,” Pattern Recognition Letters, vol. 119, pp. 3–11, 2019. [Google Scholar]

15. R. Maskeliunas, R. Damaševicius, A. Kulikajevas, E. Padervinskis, K. Pribuisis et al., “A hybrid u-lossian deep learning network for screening and evaluating Parkinson’s disease,” Applied Sciences, vol. 12, no. 22, pp. 11601, 2022. [Google Scholar]

16. A. Kouli, K. M. Torsney and W. L. Kuan, “Parkinson’s disease: Etiology, neuropathology and pathogenesis,” Exon Publications, pp. 3–26, 2018. [Google Scholar]

17. S. Bhat, U. R. Acharya, Y. Hagiwara, N. Dadmehr and H. Adeli, “Parkinson’s disease: Cause factors, measurable indicators and early diagnosis,” Computers in Biology and Medicine, vol. 102, no. 8, pp. 234–241, 2018. [Google Scholar] [PubMed]

18. T. Vaiyapuri, A. Binbusayyis and V. Varadarajan, “Security, privacy and trust in iomt enabled smart healthcare system: A systematic review of current and future trends,” International Journal of Advanced Computer Science and Applications, vol. 12, no. 2, 2021. [Google Scholar]

19. T. Vyas, R. Yadav, C. Solanki, R. Darji, S. Desai et al., “Deep learning-based scheme to diagnose Parkinson’s disease,” Expert Systems, vol. 39, no. 3, pp. e12739, 2022. [Google Scholar]

20. E. Issa, T. Salloum, B. Panossian, D. Ayoub, E. Abboud et al., “Genome mining and comparative analysis of streptococcus intermedius causing brain abscess in a child,” Pathogens, vol. 8, no. 1, pp. 22, 2019. [Google Scholar] [PubMed]

21. J. S. Stephan and S. F. Sleiman, “Exercise factors as potential mediators of cognitive rehabilitation following traumatic brain injury,” Current Opinion in Neurology, vol. 32, no. 6, pp. 808–814, 2019. [Google Scholar] [PubMed]

22. R. Aoun, H. Rawal, H. Attarian and A. Sahni, “Impact of traumatic brain injury on sleep: An overview,” Nature and Science of Sleep, vol. 11, pp. 131, 2019. [Google Scholar] [PubMed]

23. C. Laganas, D. Iakovakis, S. Hadjidimitriou, V. Charisis, S. B. Dias et al., “Parkinson’s disease detection based on running speech data from phone calls,” IEEE Transactions on Biomedical Engineering, vol. 69, no. 5, pp. 1573–1584, 2021. [Google Scholar]

24. R. Aoun, M. Kaul and A. Sahni, “Excessive daytime sleepiness due to brain tumor,” Journal of Clinical Sleep Medicine, vol. 16, no. 12, pp. 2117–2119, 2020. [Google Scholar] [PubMed]

25. G. Bar-Sela, M. J. Schultz, K. Elshamy, M. Rassouli, E. Ben-Arye et al., “Human development index and its association with staff spiritual care provision: A middle eastern oncology study,” Supportive Care in Cancer, vol. 27, no. 9, pp. 3601–3610, 2019. [Google Scholar] [PubMed]

26. J. -F. Laes, P. Aftimos, P. Barthelemy, J. Bellmunt, G. Berchem et al., “The clinical impact of using complex molecular profiling strategies in routine oncology practice,” Oncotarget, vol. 9, no. 29, pp. 20282, 2018. [Google Scholar] [PubMed]

27. M. A. Khan, M. Alhaisoni, M. Nazir, A. Alqahtani, A. Binbusayyis et al., “A healthcare system for covid19 classification using multi-type classical features selection,” Computers, Materials & Continua, vol. 74, no. 1, pp. 1393–1412, 2023. [Google Scholar]

28. B. M. Bot, C. Suver, E. C. Neto, M. Kellen, A. Klein et al., “The mpower study, Parkinson disease mobile data collected using researchkit,” Scientific Data, vol. 3, no. 1, pp. 1–9, 2016. [Google Scholar]

29. M. A. Khan, A. Khan, M. Alhaisoni, A. Alqahtani, S. Alsubai et al., “Multimodal brain tumor detection and classification using deep saliency map and improved dragonfly optimization algorithm,” International Journal of Imaging Systems and Technology, vol. 33, 2022. [Google Scholar]

30. A. R. Javed, F. Shahzad, S. ur Rehman, Y. B. Zikria, I. Razzak et al., “Future smart cities requirements, emerging technologies, applications, challenges and future aspects,” Cities, vol. 129, pp. 103794, 2022. [Google Scholar]

31. M. U. Sarwar and A. R. Javed, “Collaborative health care plan through crowdsource data using ambient application,” in 2019 22nd Int. Multitopic Conf. (INMIC), Islamabad, IEEE, pp. 1–6, 2019. [Google Scholar]

32. T. Shakeel, S. Habib, W. Boulila, A. Koubaa, A. R. Javed et al., “A survey on covid-19 impact in the healthcare domain: Worldwide market implementation, applications, security and privacy issues, challenges and future prospects,” Complex & Intelligent Systems, vol. 9, pp. 1–32, 2022. [Google Scholar]

33. K. Saleem, M. Saleem, R. Zeeshan, A. R. Javed, M. Alazab et al., “Situation-aware bdi reasoning to detect early symptoms of covid 19 using smartwatch,” IEEE Sensors Journal, vol. 23, no. 2, pp. 898–905, 2022. [Google Scholar] [PubMed]

34. C. Dhasarathan, M. K. Hasan, S. Islam, S. Abdullah, U. A. Mokhtar et al., “Covid-19 health data analysis and personal data preserving: A homomorphic privacy enforcement approach,” Computer Communications, 2022. [Google Scholar]

35. A. Amanat, M. Rizwan, M. Abdelhaq, R. Alsaqour, S. Pandya et al., “Deep learning for depression detection from textual data,” Electronics, vol. 11, no. 5, pp. 676, 2022. [Google Scholar]

36. M. Rizwan, A. Shabbir, M. Shabbir, T. Baker and D. A. -J. Obe, “Brain tumor and glioma grade classification using gaussian convolutional neural network,” IEEE Access, vol. 10, pp. 29731–29740, 2022. [Google Scholar]

37. M. Rizwan, A. Shabbir, G. Srivastava, T. R. Gadekallu, M. Shabir et al., “Risk monitoring strategy for the confidentiality of healthcare information,” Computers and Electrical Engineering, vol. 100, no. 1, pp. 107833, 2022. [Google Scholar]

38. L. H. Iwaya, A. Ahmad and M. A. Babar, “Security and privacy for mhealth and uhealth systems: A systematic mapping study,” IEEE Access, vol. 8, pp. 150081–150112, 2020. [Google Scholar]

39. A. R. Javed, M. U. Sarwar, H. U. Khan, Y. D. Al-Otaibi, W. S. Alnumay et al., “Pp-spa: Privacy preserved smartphone-based personal assistant to improve routine life functioning of cognitive impaired individuals,” Neural Processing Letters, pp. 1–18, 2021. [Google Scholar]

40. J. Li, Y. Meng, L. Ma, S. Du, H. Zhu et al., “A federated learning based privacy-preserving smart healthcare system,” IEEE Transactions on Industrial Informatics, vol. 18, no. 3, pp. 2021–2031, 2021. [Google Scholar]

41. M. Akter, N. Moustafa and T. Lynar, “Edge intelligence-based privacy protection framework for IoT-based smart healthcare systems,” in IEEE INFOCOM 2022-IEEE Conf. on Computer Communications Workshops (INFOCOM WKSHPS), New York, IEEE, pp. 1–8, 2022. [Google Scholar]

42. A. Rehman, I. Razzak and G. Xu, “Federated learning for privacy preservation of healthcare data from smartphone-based side-channel attacks,” IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 2, pp. 684–690, 2022. [Google Scholar]

43. Z. Lian, W. Wang, H. Huang and C. Su, “Layer-based communication-efficient federated learning with privacy preservation,” IEICE Transactions on Information and Systems, vol. 105, no. 2, pp. 256–263, 2022. [Google Scholar]

44. Z. Lian, Q. Yang, W. Wang, Q. Zeng, M. Alazab et al., “Deep-fel: Decentralized, efficient and privacy-enhanced federated edge learning for healthcare cyber physical systems,” IEEE Transactions on Network Science and Engineering, vol. 9, no. 5, pp. 3558–3569, 2022. [Google Scholar]

45. J. Song, W. Wang, T. R. Gadekallu, J. Cao and Y. Liu, “Eppda: An efficient privacy-preserving data aggregation federated learning scheme,” IEEE Transactions on Network Science and Engineering, pp. 1, 2022. [Google Scholar]

46. D. R. Kandati and T. R. Gadekallu, “Genetic clustered federated learning for covid-19 detection,” Electronics, vol. 11, no. 17, pp. 2714, 2022. [Google Scholar]

47. M. T. Guimaraes, A. G. Medeiros, J. S. Almeida, M. F. y Martin, R. Damaševicius et al., “An optimized approach to huntingtonâs disease detecting via audio signals processing with dimensionality reduction,” in 2020 Int. Joint Conf. on Neural Networks (IJCNN), Glasgow, IEEE, pp. 1–8, 2020. [Google Scholar]

48. A. Lauraitis, R. Maskeliunas, R. Damaševicius and T. Krilavicius, “Detection of speech impairments using cepstrum, auditory spectrogram and wavelet time scattering domain features,” IEEE Access, vol. 8, pp. 96162–96172, 2020. [Google Scholar]

49. D. Połap, M. Wozniak, R. Damaševicius and R. Maskeliunas, “Bio-inspired voice evaluation mechanism,” Applied Soft Computing, vol. 80, pp. 342–357, 2019. [Google Scholar]

50. H. Zhang, C. Song, A. Wang, C. Xu, D. Li et al., “Pdvocal: Towards privacy-preserving Parkinson’s disease detection using non-speech body sounds,” in 25th Annual Int. Conf. on Mobile Computing and Networking, pp. 1–16, 2019. [Google Scholar]

51. A. Zhan, S. Mohan, C. Tarolli, R. B. Schneider, J. L. Adams et al., “Using smartphones and machine learning to quantify Parkinson disease severity: The mobile Parkinson disease score,” JAMA Neurology, vol. 75, no. 7, pp. 876–880, 2018. [Google Scholar] [PubMed]

52. S. Grover, S. Bhartia, Akshama, Y. Abhilasha, K. R. Seeja et al., “Predicting severity of Parkinsonâs disease using deep learning,” Procedia Computer Science, vol. 132, pp. 1788–1794, 2018. [Google Scholar]

53. P. Schwab and W. Karlen, “Phonemd: Learning to diagnose Parkinsonâs disease from smartphone data,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, pp. 1118–1125, 2019. [Google Scholar]

54. J. Prince, F. Andreotti and M. De Vos, “Multi-source ensemble learning for the remote prediction of Parkinson’s disease in the presence of source-wise missing data,” IEEE Transactions on Biomedical Engineering, vol. 66, no. 5, pp. 1402–1411, 2018. [Google Scholar] [PubMed]

55. H. Zhang, A. Wang, D. Li and W. Xu, “Deepvoice: A voiceprint-based mobile health framework for Parkinson’s disease identification,” in 2018 IEEE EMBS Int. Conf. on Biomedical & Health Informatics (BHI), Las Vegas, IEEE, pp. 214–217, 2018. [Google Scholar]

56. W. Li, F. Milletarì, D. Xu, N. Rieke, J. Hancox et al., “Privacy-preserving federated brain tumor segmentation,” in Int. Workshop on Machine Learning in Medical Imaging, Springer, pp. 133–141, 2019. [Google Scholar]

57. T. M. Ali, A. Nawaz, A. U. Rehman, R. Z. Ahmad, A. R. Javed et al., “A sequential machine learning-cum-attention mechanism for effective segmentation of brain tumor,” Frontiers in Oncology, vol. 12, pp. 1993, 2022. [Google Scholar]

58. D. Stripelis, U. Gupta, H. Saleem, N. Dhinagar, T. Ghai et al., “Secure federated learning for neuroimaging,” arXiv preprint arXiv: 2205.05249, 2022. [Google Scholar]

59. M. Little, P. Mcsharry, S. Roberts, D. Costello and I. Moroz, “Exploiting nonlinear recurrence and fractal scaling properties for voice disorder detection,” Nature Precedings, pp. 1, 2007. [Google Scholar]